A learning decentralized sequential detection method based on neurocomputation

This paper proposes a neural-network learning method for a. decentralized sequential detection problem, in which each subsystem of the overall decentralized system is realized with a neural network. Suitable network models that can best match the detection problem of each subsystem are obtained. Then a reinforcement learning method is proposed to train the network fusion center and the error back propagation algorithm is used to train the network local detectors. It is shown that each neural-network subsystem can approach its ideal target - the posterior conditional probability function, in terms of minimum mean squared-error even though target values of the probability functions is not given to the learning system. This learning detection method has the advantage that it does not require statistical knowledge of observation sources and can be adapted to slowly changed environment. Simulation examples conducted on i.i.d. Gaussian data are presented for illustration.